multi-head design

[syusukee]

hi，if I would like to use multi-head for the keypoint prediction, there is only one key word("keypoint_head") in the model config, how can I realize it? Could you give me some advice? Hope for your reply! @jin-s13

[jin-s13]

Three steps to achieve this.

1. In the config, simply modify this
   

   mmpose/configs/top_down/resnet/coco/res50_coco_256x192.py

   Line 44 in 3028b9f

   keypoint_head=dict(

   
   to make keypoint_head a list of dict, for example:
   keypoint_head = [dict(...), dict(...), ...]

2. You may also modify this line, and use a for-loop to build multiple heads.
   

   mmpose/mmpose/models/detectors/top_down.py

   Line 44 in 3028b9f

   self.keypoint_head = builder.build_head(keypoint_head)

3. And modify the forward functions to support multiple heads.

https://github.com/open-mmlab/mmpose/blob/master/mmpose/models/detectors/top_down.py

================================================================================

By the way, what are these heads for? Maybe we can give an example of using multi_head models later.

[syusukee]

Thanks a lot!! I've read some paper that they combine classification or some other tasks with keypoints, so maybe you could add an example for the multi-head usage, it's a useful one!!